Building an open-source control plane for self-hosted vLLM, what would you want in it?

Every time I self-host a model I rebuild the same stuff: start the container, set up a route, check why it died overnight, remember to shut the GPU off before it burns money.

So I'm building a panel that handles it. Start/stop models, OpenAI-compatible endpoint, health checks with auto-restart, basic usage and latency stats. No Kubernetes. Works with local GPUs or rented ones (RunPod, Vast.ai). MIT licensed and self-hostable.

Not released yet, so I'd rather ask now than build the wrong thing:

- What breaks for you when self-hosting inference?

- Do you swap models often, or pin one and leave it running?

- What would you need before letting it manage a GPU unattended?

Happy to hear "just use X instead" if this already exists.